mkdir /data mkdir /models

gsutil
-m copy
-r gs://plasmic-training/prep-video-train/*
/data

docker build
-t pix-to-pix
.

gcloud docker -- pull us.gcr.io/plasmic-artefacts-2/pix-to-pix

Inference

docker run
-v $(PWD)/data/speech-to-landmarks/inference:/data/landmarks
-v $(PWD)/data/inference:/data/inferred
-v $(PWD)/models:/models
-v $(PWD)/pix-to-pix:/app
-it us.gcr.io/plasmic/pix-to-pix \

Train

nvidia-docker run
--shm-size 2G
-v /data:/data
-v /models:/models
-it us.gcr.io/plasmic-artefacts-2/pix-to-pix
bash

TODO: experiment with different settings in "train.py".

See https://github.com/NVIDIA/pix2pixHD/blob/5a2c87201c5957e2bf51d79b8acddb9cc1920b26/options/base_options.py#L13-L61

--continue_train \

python train.py
--checkpoints_dir /models
--dataroot /data
--label_nc 0
--name nius
--no_flip
--no_instance
--resize_or_crop none
--verbose